Camp & Mar/u wrrhApp/r Pnnred,n Ihe U S A
Vol
No\ I-3
II
pp 11%3lY
19X5
A NUMERICAL PROCEDURE FOR THE POROUS MEDIA EQUATION-t R C MACCAMY and EDUARDO SOCOLOVSKY Department
of Mdthemdtlcs,
Camegle-Mellon Commumcated
Umverslty,
by Matthew
Pittsburgh
PA 15213
U S A
Wttten
INTRODUCTION
In this paper we contmue a study, mltlated m [ 11, of numencal problem, u,(u2L
~(0, x) = U,,(X), u,,(x) 2 0,
procedures t > 0,
for the porous media
x E R
(P)
It 1s well known ([2]-[6]) that (P) has a umque solution u which, for each t, is supported m a fimte interval [c_(f), t+(t)] Inside the support u 1s a classical solution of (P), but across tf, U, 1s generally dlscontmuous In fact, one has
Problem (P) appears to be parabohc but it has some dlstmctly hyperbohc features When u 1s zero (P), degenerates to the hyperbohc equation u, = 2u,u, This reflects m the propagation of disturbances at fimte speed There exist several numerical methods[7-91 for the solution of(P), all of which encounter some difficulty m accurately trackmg the free boundanes In [l] a new approach to (P) was formulated with the primary purpose bemg to faclhtate front trackmg We review this approach briefly Suppose u IS a solution of (P) Define X(t, p) by the equation
X,(t, P) = -2u,(r, x0, p)),
X(0, p) = p
(1 2)
with t > 0 and p m the support of u0 Now put
U(tt PI = UO,X0, p))
(1 3)
Then one has, by (1 3), (1 2) and (P),,
5 cuxp,
= 4x,
=
+ UJX& +
UJp -
2utx,
-
ux,, 2uuJp
= 0
Hence,
w, P)XpU, P) From (1 3) we have U,(t, p) = u,(r, X0, p))X,(t,
x,u,P) The Idea of Ref p+Thts
work was supported
= -2X,‘@, P)t$k
= u,(p) p) and hence we can rewnte PI,
1 1s to reverse the above process by the National
Sctence Foundation 315
(1 4)
X(0, p) = p
(I 2) as
(1 5)
Thus one solves (1 4) and (1 5) for X
under Grant No
MCS-8001944
316
R C MACC~MY
and E SOCOLOLSKY
and U Then one defines P(t, x) and u(t, x) by x = X(t, P(r, r)),
u(r, r) = U(r, P(t, r))
(I 6)
We have not been able to give a direct proof of existence of solutions of (1 4) and ( I 5) although It appears to be possible to mfer it from the existence of weak solutions of (P) What 1s shown m [l] IS this Suppose u. satisfies ub’(x) < 0, Then, as long as a solution
-u
fu
(I 7)
(X, U) of (1 4) and (1 5) exists, it satlsfles
p) 2 1
qt, Thus the transformation
r %
I
(1 6) is possible
(1 8)
and one can then show that u, so defined.
satlsfles
(P> The important feature of the above procedure support of L&J,say a _ I x 5 a, Moreover,
is that p always lies m a fixed interval,
C-r(t) = x0, a+)
the
(1 9)
In Ref 1 we presented a very simple difference procedure based on (1 4) and (1 5) That procedure was found to track the free boundanes very accurately and easily In this paper we give an alternative vanatlonal procedure The main Interest m this second procedure IS that the coordinate change described above appears to be capable of extension to more dlmenslons and m that setting the variational treatment seems much more natural 2
We eliminate
THE
VARIATIONAL
FORMULATION
U between (I 4) and (1 5) to obtam an equation
x, = -2x;’
;
(UJX,‘),
for X
X(0, p) = p
It IS shown m Ref 1 that there exist dt most one solution of (2 I) Note that this statement Implies we need no boundary condltlons We restnct ourselves to the case m which the support of u. IS [-a, +a] and u. symmetric, uo(x) = uo( -x) Then X ~111 be antlsymmetnc with
xr = -Vr,’ $ (UO&‘), We give a vanational
formulation
Then If v(p) 1s a C”’ function
I’ 0
vanishing
X,v dp =
X(0, p) = p, of (P’)
X(r, 0) = 0,
O
(P’)
Put,
at p = 0 we obtain from (P’) and Integration
’ ((T(X~, pW(p) I0
-
by parts
W,,pMp)Q, (VP)
" I0
(x(0, p) - p)v(p)Q
The absence of a boundary condltlon of (VP) with X(r, ) E I???,
= 0
for X at p = 0 mdlcates that we should seek a solution
fi’: = {v E H,(O, u), v(0) = 0)
(2 3)
A numerul
317
procedure
Our next step IS to give an approximate vanatlonal problem Let 3; denote the subspace of @ consistmg of pIecewise lmear functions on a umform mesh of spacmg h = alN For functions m 36 the mmal condltlon is exact so the approximate vanatlonal problem IS to fmd Xh E C”‘([O, =Q)$j) such that for any vh E $,
X;vhdp =
- p(X;, p,v,gIQ = 0,
” 10(X;, PM I0 X”(O,
(AW
PI
=
p
In order to implement (AVP), for a given h, we introduce 3: with q,(zh) = 6,, and wnte
a basis q,,
J
=
,
1,
(2 4)
X%3 P) = i X,(r)(o,(p) ,=I (p, and X, of course depend on h) Then (AVP) IS equivalent
, N,
1=1,
N for
to
03
X,(O) = Jh
The mass matnx on the left of (E) is well known and given by h
a
I 0
rp,(o, dp = $z forJ
= 1,
;for/
v),v), dp = $
forJ
= I,
;forJ
dp = $
forJ
= N,
iforJ=N-
h
”
0 otherwise,
= 2,
= z 2 1, (z = 2,
,N -
1)
0 otherwise,
(2 5)
I 0 0 qNcp,
1,
0 otherwise
I 0
The right side of (E) can also be considerably
simplified
From (2 4),
O
X#, (1 -
PI = ; (X,(t)
I)h
-
X,-,(t)),
(2 6)
I = 2,3,
,N
Similarly,
$ (I %‘(P) v;(p)
= = i
1)h < p < zh,
-;
zh
< (I + l)h,
(N -
.N -
z = 1,
0 otherwlse
1 (2 7)
1)h < p < Nh,
0 otherwise
Thus the right side of (E) becomes h?a,X;’
+
h2P,(X2
h’a,(X, - X,+,)-’ h’ab(X,,
-
Xy_,)-Z
- X,)-’ + h’P,(X,+,
for z = 1, - x,)-?
for
z
=
2,
for z = N,
N
1,
(2 8)
318
R
C
MAC&MY
and E SOCOLOLSKI
where
/?, =
-f
j"+'"' u. dp +
14,(h)
th
In the next section we descnbe an lmphclt time difference scheme for (E) and give some numerical results These results Indicate convergence as h tends to zero
3
We put Xh = (X,,
NUMERICAL
PROCEDURES
, X,) and wnte (E) m the form
MXO) = f(X(t)),
x(o)
= x0
(3 1)
Here the matnx M 1s determmed by (2 5) and the nonlinear function f IS given by (2 8) We give an lmphclt time difference scheme for (3 1) Choose d t and define the sequence x”, n = 0, 1, 2, , by the formulas MXn+l
-
Arf(X”+l)
MX” = 0,
-
n 2 0,
As At tends to zero the x” will converge to X(nAt) Equation (3 2) can be solved by Newton Iteration F, RN --, RN by F,(Y)
= MY
- Atf(Y)
-
X0 = X0
For each
(3 2)
n define
a function
MX”
(3 3)
so that (3 2) IS simply F,(X”+‘) = 0 Note that only the term MX” changes with n m F,, For each n 2 0 we define sequences X;, k = 0, 1, 2, , by the formulas l’,F(X;+‘)X;;:
= X;” kz0,
-
F(X;+’
X;;”
= X”
-
F(X;-‘), (3 4)
In order to check the convergence of our method we made some numerlcal on the exphclt solution of (P) given m Ref IO That solutlon IS the dlstrlbutlon a delta function at the ongm It has the form
- &)
qt, x) =
4 (9t)-1’3
for 1x1S (9t)“’
(3 5)
elsewhere
0
i
experiments produced by
We consldered then the function i(t + I, n) so that at f = 0 the support of u0 IS [ - 9”3, 9’ 3] We first camed out our numerical procedure with eleven nodes that IS h = 9”‘/ 10 = 0 208 Then we used the time dlfferencmg with At = h In Table 1 we present the results for the Table
OOiIOO 0 4160 0 X320 I 2481 I 6641 2 0801 2 4961
2 2 2 2 2 3 3
080.084 335,807 545,220 724,893 883,562 026 461 157,007
(h = 0 208,
ZA it
r:
t
I
2 2 2 2 2 2 3
080,084 322,060 525 132 70 1,345 857 979 999,639 129,422
N =
IO) rel
error 0000000 0 013 747 0 020 098 0 023,548 0 025 583 0 026,822 0 027 575
_
error
= 0 0 0 0 0 0 0
Ci’_ -
000,000 005 885 007 896 008,654 008 872 008 863 008,734
A numerical Table2
319
procedure
r = 12481
P+(f)
N 2 2 2 2
10 40 160 640
724 893 724 893 724,893 724 893
error 2 2 2 2
701 345 719,166 723,472 724,539
0 0 0 0
023,548 005,727 001,421 000,354
rel error 0 0 0 0
008,654 002,101 000,522 000,130
position
of the nght-hand free boundary c+ at several times (“, and (“, represent the exact and approximate values respectively We also give absolute and relative errors We see that the free surface IS fairly accurately tracked even with a coarse grid In order to determine convergence rates we did the calculations with four different h values correspondmg to N = 10, 40, 160 and 640 We found the L, errors, defined as the sum of the errors at the gnd points times h to be almost exactly of order h As a typical calculation we present in Table 2 the exact and approximate values of (+ at the time 1 248 1 of Table 1 for the four N values Table 2 indicates that the error m the free surface, like the L, error, 1s of order H The difference scheme m Ref 1 appeared expenmentally to be of order h2 but It required At = II”’ so the two methods are competltlve We performed some further numerical experiments with exphclt time differences We encountered stability problems with this method unless we chose At = h3, reflecting the nonlinear nature of the problem Thus explicit differences would appear to be rather unsatisfactory for this problem
REFERENCES 1
2 3 4 5 6 J 8 9 10
M E Gurtm, R C Mac&my and E Socolovsky, A coordmate transformation for the porous media equation that renders the free-boundary stationary Quarr of Appl Math 47, 345-358 (1984) 0 A Oletmk, A S Kahshmkov and Y-L Chzou, The Cauchy problem and boundary value problems for equations of the type of non-stattonary filtratton Izv Akd Nauk SSR Ser Mat 22, 667-704 (1958) A S Kahshmkov, On the occurrence of smgulanttes m the solutions of the equation of nonstatlonary filtratton, Vychd Mat Mat FIS 7 440-444 (1967) D G Aronson, Regulanty propertIes of flows through porous media SIAM J Appl Moth 17, 461-468 (1969) D G Aronson, Regulanty properues of flows through porous media the Interface Arch Rnr Mech And 37, l-10 (1970) B F Knerr, The porous medmm equation m one-dlmenslon Truns Amer Math Sot 234, 381-415 (1977) E DlBenedetto and D C Hoff, An Interface trackmg algorithm for the porous medmm equatton Math Res Center, Uruv of Wlsconsm, Tech Summary Report 1489 (1983) J L Graveleau and P Jamet, A finite dtfference approach to some degenerate nonlinear parabohc equations SIAM J Appl Math 20, 199-223 (1971) M Mlmura and K Tomolda, Numerical approxlmatlons for Interface curves to a porous media equation (to appear) R E Pattle, Dlffuslon from an mstantaneous pomt source with a concentratton dependent coefficient Quorf J Mech Appl Math 12, 407-409 (1959)